​

Peak Guard is a real-time dynamics control and monitoring system designed to stabilise transient-heavy audio while maintaining low-latency behaviour suitable for interactive use cases such as film, video game, broadcast ect.
 

• Real-time Windows audio processor (WASAPI)

• Low-latency dynamics control (no lookahead)

• Deterministic processing with safety ceiling

• Built for interactive audio (games, broadcast, live)
   

At its core, the AudioProcessor implements a sample-by-sample peak-driven control pipeline operating on interleaved floating-point buffers. For each frame, it derives a per-frame peak across channels which feeds a shared envelope detector. This detector uses asymmetric attack/release coefficients (exponential smoothing) to track signal amplitude with fast attack and controlled release characteristics, ensuring transient responsiveness without excessive modulation.
 

The processing chain is structured in three stages:

• Compressor stage
A threshold/ratio model is applied in the linear domain. When the tracked envelope exceeds the threshold, a target level is computed and converted into a gain factor. This gain is smoothed implicitly via the envelope follower rather than a separate gain smoother, reducing state complexity and avoiding additional latency.

• Limiter stage (pre-output gain)
A peak limiter operates on the post-compressor signal. Attack is effectively instantaneous, while release is time-based, preventing sustained gain reduction after transient events. This stage ensures that downstream gain increases do not reintroduce overs.

• Post-gain safety ceiling

A final non-user-facing safety limiter operates after output gain. This stage enforces a fixed ceiling (~-0.8 dBFS) and protects against digital clipping regardless of user-applied gain. It uses a fast peak-tracking envelope and applies corrective gain only when required, remaining transparent under normal conditions.
 

The combined gain structure can be expressed as:

g_total = g_compressor × g_limiter × g_output × g_safety

All stages operate per-sample with no lookahead, prioritising determinism and minimal latency over brickwall-style peak anticipation.
 

To support observability, the processor accumulates statistical telemetry during processing:

• RMS (input/output) computed from running sum-of-squares
• Peak output tracking and exact clip counting
• Gain reduction decomposition:

• Compressor GR

• Limiter GR

• Safety limiter GR

• Combined GR
   

To reduce CPU overhead at low buffer sizes, meter calculations are decimated (e.g. every other block), while the audio path remains unaffected. This avoids UI-induced performance spikes in ultra-low-latency configurations.

The AudioDeviceManager manages system-level routing and stream lifecycle. It interfaces with WASAPI in shared mode for broad compatibility, with optional VB-Cable routing for full system audio capture. Device enumeration is performed via IMMDeviceEnumerator, and default endpoint switching is handled through IPolicyConfig for Console, Multimedia, and Communications roles.
 

A fallback WASAPI path is provided to allow direct device processing without virtual routing dependencies, improving robustness in constrained or unsupported environments.
 

Latency is controlled through a configurable target buffer model. The system attempts progressively lower buffer sizes during initialisation, falling back gracefully if stability constraints are encountered. A bounded FIFO is used to decouple capture and render timing, with overwrite-on-full behaviour to prevent latency accumulation under load.

On the UI side, Peak Guard exposes real-time system state through a set of time-series visualisations:

• Gain Reduction over time (including safety stage contribution)
• Latency behaviour and buffer performance
• Input/output level monitoring
 

These plots are driven by timestamped telemetry streams and rendered in a sliding window, enabling users to correlate audible behaviour with system response in real time.
 

The overall design prioritises:

• Deterministic, low-latency processing (no lookahead, minimal buffering)
• Clear separation between audio path and telemetry
• Fail-safe behaviour under misconfiguration (safety ceiling)
• System-level integration with explicit device and routing control
 

This results in a dynamics control system that is not only corrective, but also inspectable, allowing users to understand how and when processing is applied, rather than treating it as a black box.